Clustering and Unsupervised Classification
The classification techniques treated in Chap. 8 all require the availability of labelled training data with which the parameters of the respective class models are estimated. As a result, they are called supervised techniques because, in a sense, the analyst supervises an algorithm’s learning about those parameters. Sometimes labelled training data is not available and yet it would still be of interest to convert remote sensing image data into a thematic map of labels. Such an approach is called unsupervised classification since the analyst, in principle, takes no part in an algorithm’s learning process. Several methods are available for unsupervised learning. Perhaps the most common in remote sensing is based on the use of clustering algorithms, which seek to identify pixels in an image that are spectrally similar. That is one of the applications of clustering treated in this chapter.